|其他公開專利號||CN1081359C, CN1148207A, EP0745933A2, EP0745933A3|
|公開號||08459490, 459490, US 5644780 A, US 5644780A, US-A-5644780, US5644780 A, US5644780A|
|發明人||David Arnold Luick|
|原專利權人||International Business Machines Corporation|
|匯出書目資料||BiBTeX, EndNote, RefMan|
|專利引用 (14), 被以下專利引用 (19), 分類 (18), 法律事件 (7)|
|外部連結: 美國專利商標局, 美國專利商標局專利轉讓訊息, 歐洲專利局|
The present invention relates to a random access read/write memory (RAM) device with multiple read and write ports or a multiple port register file, and more particularly to a high speed register file including 64 or more ports and adapted for use with Very Long Word Instruction (VLIW) and N-way superscaler processors.
Known multiple or multi-port register files are inadequate for use with Very Long Word Instruction (VLIW) and N-way superscaler processors. VLIW processors consist of a series of arithmetic logic or functional units that can execute parts of instructions simultaneously. VLIW and N-way superscaler processors with 8, 16, or more pipeline functional units require a central register file for communication that must provide three or four ports per pipeline for perhaps a 64-port or more requirement. Current technology simply does not allow construction of a direct physical implementation of so many ports, especially with very stringent access and write time requirements. Currently the most ports physically wireable may be less than or equal to twenty (20).
A principal object of the present invention is to provide an improved multiple port register file for use with Very Long Word Instruction (VLIW) and N-way superscaler processors.
In brief, a high speed register file is provided for use with Very Long Word Instruction (VLIW) and N-way superscaler processors. The high speed register file includes a selected number of copies of a general purpose register (GPR) building block. The GPR building block includes at least two interleaved sub-banks of registers. Each of the sub-banks includes a number N of write ports and a number M of read ports. The sub-banks are interleaved by write ports and have non-interleaved read ports.
The present invention together with the above and other objects and advantages may best be understood from the following detailed description of the preferred embodiments of the invention illustrated in the drawings, wherein:
FIG. 1 is a block diagram illustrating a pair of twenty port general purpose register (GPR) file physical building blocks in accordance with the invention;
FIG. 2 is a block diagram illustrating a sixty-four port register file including four copies of the twenty-eight port logical GPR file logical building block of FIG. 1;
FIG. 3 is a block diagram illustrating a write port partitioning arrangement utilizing the GPR file building blocks of FIG. 1;
FIG. 4 is a block diagram illustrating an alternative reduced write port partitioning register file arrangement of the invention utilizing the GPR file building blocks of FIG. 1;
FIG. 5 is a block diagram illustrating a Very Long Word Instruction (VLIW) processor unit including the sixty-four port register file of FIG. 4;
FIG. 6 is a block diagram illustrating a seventy-two port register file including four copies of the twenty eight port register file building block of FIG. 1;
FIG. 7 is a schematic diagram illustrating a cell of the register file building block of FIG. 1 together with an address decode and an arithmetic and logic unit (ALU) adder;
FIG. 8 is a schematic diagram illustrating a variable performance read access in accordance with the invention of the GPR file building block of FIG. 1;
FIG. 9 is an exemplary layout of the GPR file building blocks of FIG. 1 bit 0;
FIG. 10 is another exemplary layout of the GPR file building blocks of FIG. 1 bit 0 illustrating an alternative read port layout in accordance with the invention; and
FIG. 11 is a schematic diagram illustrating an exemplary layout of a processor unit including the GPR file building blocks of FIG. 1.
Having reference now to the drawings, FIG. 1 illustrates a twenty-eight port general purpose register (GPR) file logical building block in accordance with the invention generally designated by the reference character 50. In accordance with a feature of the invention, interleaving of write ports, specifically into even and odd sub-banks of registers as seen by write ports is employed. The GPR register file building block 50 of depth 32 yields a total of 16 available write ports and 12 available read ports.
GPR file logical building block 50 includes a pair of twenty port GPR file physical building blocks 52. The twenty port GPR file physical building blocks 52 include interleaved write ports of an even GPR sub-bank 52 and an odd GPR sub-bank 52. The GPR sub-banks 52 provides 12 read ports and 8 write ports with one sub-bank of 32 even registers and 8 write ports with another sub-bank of 32 odd registers. The GPR register file building block 50 includes sixteen physical write ports and twelve physical read ports that are used to construct a 64-port GPR logical register file generally designated by 60 in FIG. 2 with excellent performance characteristics.
Referring now to FIG. 2, the sixty-four (64) port general purpose register file 60 includes four copies of the twenty-eight port register file logical building block 50 of FIG. 1. With the GPR file logical building block 50 copied four times, a 4-fold increase in read ports is provided for a required number of 48 total read ports. For the 64-deep register file 60, the 64 registers are divided into two groups; one group of 32 even and one group of 32 odd-numbered registers. In use, this register file arrangement 60 allows the compiler to generate basic blocks of instructions that have an approximately equal number of odd and even-numbered target register operands through some combination of register allocation and renaming algorithm. In the case of tightly encoded VLIW instructions, each instruction parcel has either an even (even parcel) or odd target operand register file port permanently assigned, obviating the need for the last bit of the register address to be encoded in the instruction.
Write performance is improved in the interleave register file 60 since only an 8-way write selection is required, as compared with a 16-way write selection in an uninterleaved register file, before a write is performed in the target register. However, no increase in write ports is possible from copying since all copies receive all writes to remain valid copies. The GPR 64-port register file 60 also improves the actual read port performance of the register file since it places the copy 1, 2, 3 and 4 closest to the pipeline data flow it drives. A disadvantage of the copy strategy is that target operands must physically traverse the entire area of the four copies of GPR logical building blocks 50 in the register file 60 of FIG. 2.
FIG. 3 is a block diagram illustrating a write port partitioning register file arrangement generally designated by 68 utilizing the four copies of the GPR building block 50. Each block 70 includes two GPR building blocks 52 with eight additional write paths or ports into each cell. In the register file 68, any load can write all copies 1, 2, 3 and 4 simultaneously and any ALU can write all copies 1, 2, 3 and 4 simultaneously. The arrangement 68 is impractical due to the required eight additional write ports required to provide 16 ALU write ports and 8 load write ports.
FIG. 4 illustrates an alternative reduced write port partitioning register file arrangement generally designated by 78 including four blocks 80. Each block 80 includes two 1/2 blocks or sub-banks 52 or one block 50. In copies 1 and 2, one of the sub-banks 52 is used for ALU writes pipes 0-7 and the other sub-bank 52 is used for 8 load write ports. In copies 3 and 4, one of the sub-banks 52 is used for ALU writes pipes 8-15 and the other sub-bank 52 is used for 8 load write ports. The partitioning arrangement 78 of FIG. 4 provides a practical solution requiring a total of sixteen write ports, rather than the impractical arrangement 68 of FIG. 3 that requires a total of 24 write ports. A sophisticated compiler can limit target writes only to two of the four copies 80 in a given cycle. For ALU operations producing target results that must be written to the register file 78, the compiler arranges to have all or most of the work of a basic block confined to either the first eight pipelines 0-7 or the second eight pipelines 8-15 of the sixteen total pipeline slots or pipes in a VLIW processor.
Each group of two copies 80 of the register file 78 need only receive the ALU target results from one half of the sixteen pipes; either eight pipes 0-7 or eight pipes 8-15, reducing the total number of write ports required by eight per copy. Additional read ports are not particularly expensive; however, additional global write ports are very expensive since they impact all copies 1, 2, 3 and 4 with extra input ports. When pipes 8-15 need to look at a result from pipes 0-7 and vice versa, the compiler schedules move register operations and schedules register references appropriately so that pipe stalls for this reason are totally avoided.
FIG. 5 is a block diagram illustrating a VLIW processor unit generally designated by 82 including the sixty-four port GPR file 78. Another significant benefit from the ALU target partitioning scheme illustrated in FIG. 5 is a substantial reduction in the total bus wire length required to get any ALU result to any write port, as well as reduced loading requirements, two versus, four without the write port partitioning. A line labeled 84 represents a longest ALU write bus wire required with the partitioned register file 78. A dotted line labeled 86 represents a longest ALU write bus wire required with the non-partitioned register file 60. In the partitioned ALU target register file 78, at worst, an ALU need only drive to its own and an adjacent register file copy as opposed to driving to its next adjacent copy in the non-partitioned arrangement. This also potentially allows these bus wires to be double thickness and double width wires. The thicker and wider wires have lower resistance to provide better performance and also because the bus wires may run only about one-half length as compared to a non-partitioned register file arrangement. Write port target partitioning is explicitly not used for cache data load write ports. This is so partly because in commercial environments, performance is often limited by the number of load pipes available, not the number of ALUs. Thus, ALU operations can be duplicated in each ALU target partition with little interference, loads cannot. Also, many load dependency cases exist, for example, linked lists, so that latency is quite important. There is almost no cycle time/wire delay difference for the cache data load case because the data busses are already long wires with high-current drivers. Driving four register file copies versus only two causes an insignificant change in bus delay, especially given that load paths are already two cycle paths.
FIG. 6 illustrates a seventy-two port register file generally designated by 90 including four copies of the twenty eight port GPR file logical building block 50 of FIG. 1. In FIG. 6, the completed register file layout includes eight of the basic 20-port GPR sub-banks 52 yielding a logical register file capable of a maximum of 48 reads and 24 writes per cycle for a total of 72 logical ports with very fast read access and write-through performance capability.
FIG. 7 is a schematic diagram illustrating a cell of the register file building block 52 of FIG. 1. A DCACHE data input is applied to an 8-way byte align 100 that is connected at its output to a 4-way multiplexer (MUX) 102. MUX 102 is connected to a 2-way multiplexer (MUX) 104 that together define an 8-way MUX function included in a cell 106. Cell 106 includes a pair of latches, shown as L1 and L2 latches or a shift register latch. The L1 latch is connected to a 8-way MUX 108 that is connected at its output to a 8-way MUX 110. A GPR address decoder generally designated by 112 including a pair of address buffers 114, 116 and address L1 latches 118, 120 applies address signals to the MUXs 108 and 110. A C1 clock output is used to set the address buffers 114, 116 for the L1 and L2 latches 118 and 120. The output of MUX 110 is applied to an adder 122 that is connected to an L2* latch 124 of an associated arithmetic logic unit (ALU). The output of the L2* latch 124 is applied to a 4-way MUX 126 that is connected to the 2-way MUX 104 via a global network wire labeled 128. In a complementary metal oxide semiconductor (CMOS) integrated circuit implementation with six levels of metallization, the global network wire 128 can be fifth level of metal and made thicker, and for example, twice the standard wiring width. The cell output of L2 latch 106 is applied to a 64-way MUX 130 that provides a restore path.
FIG. 8 illustrates a variable read port performance feature of the invention. For each register file copy 50, not all the twelve read ports have the same performance requirements. Eight of the read ports are required to be write through and as fast as possible. Four of the fast read ports feed the right input of the ALU and as such must have the true and complement function for subtract and other functions. The other four fast read ports feed the other ALU input (not complemented) and the CACHE logic (machine critical path) and also are made as fast as possible. An L1 latch 142 provides two latch phases, -PHASE and +PHASE. The L1 latch +PHASE is used to drive a plurality of read port selectors 144 and 146, providing fast L1 write through ports. The write through ports pass data from a write port to the read port in the same clock cycle. Also the compiler can guarantee that no more than two output read selectors will access the same register on any given cycle, limiting the current that must be provided and improving performance further. The L1 latch -PHASE is used to drive a buffer 148 that is coupled to a pair of selects 150 and 152, providing the four slow L2 read ports that are non-write though and do not have stringent performance requirements as they are used for either store register data access or to load exception save restore registers, both requiring noncritical timings. The four slow read ports can tolerate longer, thinner, and more irregular wire bussing which gives the designer an additional degree of freedom. A bypass select 154 coupled to the write port provides fast read ports to the ALU pipes.
FIG. 9 provides an exemplary layout of the GPR file logical building block 50 of FIG. 1 bit 0. Near the middle of the layout, 64 GPR register cells are shown divided into 32 odd cells 106 and 32 even cells 106. An adjacent column of 8:1 multiplexers defining 4-way and 2-way MUXs 102, 126 and 104 is connected to the even and odd cells 106 by a plurality of write lines indicated by example lines labeled 160. Eight L1 read ports and four L2 read ports are illustrated with lines indicating example connection of the cells 106 to the read ports.
Referring to FIG. 10, another exemplary layout of the GPR file logical building block 50 of FIG. 1 bit 0 illustrates an alternative read port layout in accordance with the invention. As depicted and described with respect to FIG. 8, GPR file logical building block 50 includes eight fast L1 read ports and four slow L2 read ports. Four fast L1 read ports are provided between the write decode, clocking, buffering and test and a first four eight-way write MUXs. Four fast L1 read ports are provided between a second four eight-way write MUXs and the four slow L2 read ports. Exemplary connecting lines are shown between eight write port inputs and the first four eight-way MUXs, these eight-way MUXs and the cells and the cells and one of the four slow L2 read ports.
FIG. 11 is a schematic diagram illustrating an exemplary layout of a processor unit including the GPR file 78. In FIG. 11, the partitioned GPR file 78 is shown generally centrally located above the register/register (RR) pipes 0-7 and below the register/storage (RS) pipes 0-5 and an associated DECACHE directory (DDIR) and segment look-aside buffers (SLBs) 0-7. An exemplary wire is illustrated between the even and odd GPR 1/2 copies 1 (sub-banks 52 in FIG. 1) and a decode and control area that can be provided with thicker wire, for example, two times standard width.
While the present invention has been described with reference to the details of the embodiments of the invention shown in the drawing, these details are not intended to limit the scope of the invention as claimed in the appended claims.
|US4128883 *||1977年9月30日||1978年12月5日||Ncr Corporation||Shared busy means in a common bus environment|
|US4554645 *||1983年3月10日||1985年11月19日||International Business Machines Corporation||Multi-port register implementation|
|US4558433 *||1983年5月31日||1985年12月10日||International Business Machines Corporation||Multi-port register implementations|
|US4652993 *||1984年4月2日||1987年3月24日||Sperry Corporation||Multiple output port memory storage module|
|US4766535 *||1985年12月20日||1988年8月23日||International Business Machines Corporation||High-performance multiple port memory|
|US4811296 *||1987年5月15日||1989年3月7日||Analog Devices, Inc.||Multi-port register file with flow-through of data|
|US4866603 *||1988年1月20日||1989年9月12日||Fujitsu Limited||Memory control system using a single access request for doubleword data transfers from both odd and even memory banks|
|US5111431 *||1990年11月2日||1992年5月5日||Analog Devices, Inc.||Register forwarding multi-port register file|
|US5204841 *||1992年4月23日||1993年4月20日||International Business Machines Corporation||Virtual multi-port RAM|
|US5206944 *||1990年6月7日||1993年4月27日||The United States Of America As Represented By The Secretary Of The Air Force||High speed analog to digital converter board for an IBM PC/AT|
|US5276848 *||1991年8月20日||1994年1月4日||International Business Machines Corporation||Shared two level cache including apparatus for maintaining storage consistency|
|US5289427 *||1992年7月20日||1994年2月22日||Motorola, Inc.||Multiport memory with write priority detector|
|US5481495 *||1994年4月11日||1996年1月2日||International Business Machines Corporation||Cells and read-circuits for high-performance register files|
|US5481736 *||1993年2月17日||1996年1月2日||Hughes Aircraft Company||Computer processing element having first and second functional units accessing shared memory output port on prioritized basis|
|US6219777 *||1998年7月10日||2001年4月17日||Nec Corporation||Register file having shared and local data word parts|
|US6301653||1998年10月14日||2001年10月9日||Conexant Systems, Inc.||Processor containing data path units with forwarding paths between two data path units and a unique configuration or register blocks|
|US6343356 *||1998年10月9日||2002年1月29日||Bops, Inc.||Methods and apparatus for dynamic instruction controlled reconfiguration register file with extended precision|
|US6430677||2001年2月28日||2002年8月6日||Bops, Inc.||Methods and apparatus for dynamic instruction controlled reconfigurable register file with extended precision|
|US6629232 *||2000年7月3日||2003年9月30日||Intel Corporation||Copied register files for data processors having many execution units|
|US6704855||2000年6月2日||2004年3月9日||International Business Machines Corporation||Method and apparatus for reducing encoding needs and ports to shared resources in a processor|
|US7024540 *||2003年10月28日||2006年4月4日||Pts Corporation||Methods and apparatus for establishing port priority functions in a VLIW processor|
|US7120780||2002年3月4日||2006年10月10日||International Business Machines Corporation||Method of renaming registers in register file and microprocessor thereof|
|US7366032 *||2005年11月21日||2008年4月29日||Advanced Micro Devices, Inc.||Multi-ported register cell with randomly accessible history|
|US7401205 *||2000年8月11日||2008年7月15日||Mips Technologies, Inc.||High performance RISC instruction set digital signal processor having circular buffer and looping controls|
|US7840783 *||2007年9月10日||2010年11月23日||Netlogic Microsystems, Inc.||System and method for performing a register renaming operation utilizing hardware which is capable of operating in at least two modes utilizing registers of multiple widths|
|US8533435 *||2010年9月3日||2013年9月10日||Nvidia Corporation||Reordering operands assigned to each one of read request ports concurrently accessing multibank register file to avoid bank conflict|
|US20030167388 *||2002年3月4日||2003年9月4日||International Business Machines Corporation||Method of renaming registers in register file and microprocessor thereof|
|US20040093484 *||2003年10月28日||2004年5月13日||Pts Corporation||Methods and apparatus for establishing port priority functions in a VLIW processor|
|US20070239967 *||2007年3月27日||2007年10月11日||Mips Technologies, Inc.||High-performance RISC-DSP|
|US20070239970 *||2006年4月6日||2007年10月11日||I-Tao Liao||Apparatus For Cooperative Sharing Of Operand Access Port Of A Banked Register File|
|US20110072243 *||2010年9月3日||2011年3月24日||Xiaogang Qiu||Unified Collector Structure for Multi-Bank Register File|
|USRE40883||2004年4月19日||2009年8月25日||Altera Corporation||Methods and apparatus for dynamic instruction controlled reconfigurable register file with extended precision|
|WO2003036468A1||2001年10月24日||2003年5月1日||Telefonaktiebolaget Lm Ericsson (Publ)||An arrangement and a method in processor technology|
|美國專利分類號||712/23, 712/E09.027, 711/157, 712/215, 712/E09.071, 712/24, 712/E09.026, 711/127|
|國際專利分類號||G06F9/30, G06F9/38, G06F9/34, G06F12/08|
|合作分類||G06F9/3885, G06F9/3012, G06F9/30141|
|歐洲分類號||G06F9/30R5, G06F9/30R6, G06F9/38T|
Owner name: INTERNATIONAL BUSINESS MACHINES CORPORATION, NEW Y
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:LUICK, DAVID ARNOLD;REEL/FRAME:007587/0399
Effective date: 19950712
|1997年9月2日||CC||Certificate of correction|
Year of fee payment: 4
Year of fee payment: 8
|2009年1月5日||REMI||Maintenance fee reminder mailed|
|2009年7月1日||LAPS||Lapse for failure to pay maintenance fees|
|2009年8月18日||FP||Expired due to failure to pay maintenance fee|
Effective date: 20090701